Why Does ChatGPT Suddenly ‘Think Step‑by‑Step’? Unveiling the Chain‑of‑Thought Emergence

Since its release, ChatGPT has maintained high popularity, and its impact is not only due to better dialogue performance but also because it exhibits “emergent” capabilities from a technical perspective.

The most representative emergent ability is the “Chain of Thought” (CoT), which enables step‑by‑step reasoning. An example shows GPT‑3 failing a simple math problem, but when prompted with “let’s think step by step”, it produces the correct answer and a full reasoning trace.

These abilities are not explicitly trained; they appear when researchers add prompts that trigger multi‑step reasoning, suggesting a sudden emergence.

Scaling studies across many datasets reveal a common pattern: model accuracy stays low until a certain size threshold, after which performance on tasks requiring multi‑step reasoning (e.g., math word problems) jumps sharply, while tasks solvable by fast intuition (System 1) improve smoothly.

The phenomenon aligns with Kahneman’s System 1 / System 2 theory: emergent abilities correspond to System 2‑type tasks that need deliberate, sequential reasoning.

One hypothesis links the emergence to the inclusion of code in training data, because code often contains explicit multi‑step logic, teaching models to perform step‑by‑step inference.

Although this remains a hypothesis, ongoing research aims to clarify why code data triggers the Chain‑of‑Thought capability and how it relates to human System 2 reasoning.